01/08/2026
Anthropic tested Claude for safety. Instead, Claude mistook the real internet for a capture-the-flag game and breached 3 real organizations.
The AI safety test created the very breach it was supposed to prevent.
We're not ready for autonomous AI. Not even close.